An attack that copies a model’s behavior by asking it many questions.
It is like asking a paid homework app ten thousand questions. Write down every answer, and you build your own answer key.
Attackers send huge batches through an API. They use the replies to make a copy that acts similarly.
API
Attackers use an API to collect many questions and answers.
Model weights
This attack usually copies behavior without stealing the model weights.
Anti-distillation
Anti-distillation can limit leaked outputs and make copying harder.